Type2Branch: Keystroke Biometrics based on a Dual-branch Architecture with Attention Mechanisms and Set2set Loss

TL;DR

Type2Branch tackles scalable keystroke dynamics verification for very large user populations by integrating synthetic timing features derived from a population profile, a dual-branch CNN-RNN embedding network with attention, and a novel Set2set loss within an increasing-difficulty curriculum. The Set2set Loss extends prior SetMargin concepts to multiple sets and adds a radius-penalty to regularize the embedding space, enabling robust verification under a fixed global threshold. Empirical evaluation on the KVC-onGoing desktop and mobile tasks yields state-of-the-art mean per-subject EERs of 0.77% and 1.03% with five enrollment samples, and 3.33% and 3.61% under global-threshold settings, outperforming previous methods. Cross-database tests and ablation studies confirm the gains are due to the dual-branch design and Set2set Loss, with synthetic features and curriculum contributing additional improvements; limitations include reliance on transcription-based data and zero-effort impostor evaluations. The work suggests that embedding-space regularization via Set2set and population-aware features can scale behavioral biometrics while maintaining accuracy.

Abstract

In 2021, the pioneering work TypeNet showed that keystroke dynamics verification could scale to hundreds of thousands of users with minimal performance degradation. Recently, the KVC-onGoing competition has provided an open and robust experimental protocol for evaluating keystroke dynamics verification systems of such scale. %, including considerations of algorithmic fairness. This article describes Type2Branch, the model and techniques that achieved the lowest error rates at the KVC-onGoing, in both desktop and mobile typing scenarios. The novelty aspects of the proposed Type2Branch include: i) synthesized timing features emphasizing user behavior deviation from the general population, ii) a dual-branch architecture combining recurrent and convolutional paths with various attention mechanisms, iii) a new loss function named Set2set that captures the global structure of the embedding space, and iv) a training curriculum of increasing difficulty. Considering five enrollment samples per subject of approximately 50 characters typed, the proposed Type2Branch achieves state-of-the-art performance with mean per-subject Equal Error Rates (EERs) of 0.77% and 1.03% on evaluation sets of respectively 15,000 and 5,000 subjects for desktop and mobile scenarios. With a fixed global threshold for all subjects, the EERs are respectively 3.25% and 3.61% for desktop and mobile scenarios, outperforming previous approaches by a significant margin. The source code for dataset generation, model, and training process is publicly available.

Figures (6)

• Figure 1: Overview of key aspects of Type2Branch. The novel loss function, along with the proposed input features, training curriculum, and dual-branch architecture, make up the proposed biometric keystroke verification system.
• Figure 2: Representation of the feature extraction process. The key code sequence is used to generate synthetic timings based on the general population profile, which are considered together with the original keystroke timings.
• Figure 3: Type2Branch: proposed dual-branch (recurrent and convolutional) embedding model for distance metric learning.
• Figure 4: t-SNE projections of embeddings for 20 subjects, comparing the proposed Set2set Loss with SetMargin Loss and Triplet Loss. Both the variance of the average class radius and the overlap between classes are minimized for the Set2set Loss.
• Figure 5: Effect on the EER with a global threshold when increasing $K$ (number of sets) in the loss function. We compare a reduced model (left y-axis) and the full model (right y-axis). Given enough training data, a larger $K$ improves the performance by providing a more comprehensive purview of the embedding space.